Documentation

Using Amanu

Hold a key, speak, release — your words appear at the cursor in whatever app you are in, transcribed entirely on your Mac.

Overview

Amanu is a push-to-talk dictation app for macOS. Put your cursor anywhere, hold a key while you talk, and let go — your words land where you were typing, in the app you are already in. Speech recognition and the optional cleanup pass both run on your Mac. There is no account, no cloud, and no audio leaving the machine.

The name is from the Latin a manu, “at hand”: the Roman amanuensis who stood beside you and wrote down what you spoke.

  • Instant — words land at the cursor the moment you release the key.
  • Every app — native insertion where it is supported, paste-swap where it is not.
  • On-device — transcription and cleanup never touch the network.
  • Private by construction — no audio leaves your Mac, by design rather than by policy.

Requirements

  • Apple silicon Mac (M1 or later).
  • macOS 15.0 or later.
  • Around 600 MB of disk for the speech model, plus optional space if you add cleanup models.
  • Dictation in 28 European languages; Refine cleanup in English and seven more. See Languages.
  • One app process — no Python, no helper daemons, no embedded interpreters.

First run

On first launch Amanu walks you through a short setup. You can revisit any of it later in Settings.

  1. Microphone — grant access so Amanu can hear you while you hold the key.
  2. Accessibility — needed for the global hotkey, for inserting text at the cursor, and for detecting secure fields. See Privacy and history for exactly what this is used for.
  3. Model download — the speech model (about 600 MB) downloads once, over HTTPS, checksum-verified.
  4. Try it — a built-in field to dictate your first sentence before you finish onboarding.
Set up Amanu Step 1 of 4

Microphone access

Amanu needs the microphone to hear you. Audio is transcribed on device and never leaves your Mac.

The first-run setup, rebuilt from the app. Buttons here are illustrative; in Amanu they trigger the real macOS prompts.

The hotkey

By default, hold fn — the globe key — to talk, and release to insert. That is the whole interface, with nothing to configure.

  • Rebind to any key or combination in Settings.
  • Hold or toggle — hold-to-talk (push-to-talk, like a radio) or toggle mode (press once to start, again to stop).
  • Debounce — tune how long a press must last so an accidental tap is ignored.

Dictating

  1. Put your cursor where you want the text.
  2. Hold your hotkey and speak.
  3. Release. Amanu transcribes on the Neural Engine and inserts the text at the cursor.

While you hold the key, a small HUD capsule shows a live waveform. It is a non-activating panel, so it never takes focus and cannot steal your cursor. When transcription finishes it shows the outcome — inserted, canceled, or an error — and then leaves. The HUD honors Reduce Motion and announces its state to VoiceOver.

Voice commands

A few phrases are interpreted as commands rather than transcribed literally. They are parsed deterministically — never guessed by a model — so they behave the same way every time.

  • new line — inserts a line break.
  • new paragraph — inserts a blank line.
  • scratch that — removes the last thing you dictated.
  • all caps <word> — uppercases the next word.

Languages

Amanu dictates in 28 European languages — Latin, Cyrillic, and Greek scripts — from one setting: Settings > Transcription > Dictation language. English is the default. The setting is a hint that sharpens recognition for the language you speak; the same on-device model handles all of them, so switching languages downloads nothing.

Refine’s Punctuate and Clean modes run for English, Spanish, French, German, Italian, Portuguese, Russian, and Polish — always in the language you spoke, never translating. For any other dictation language the cleanup pass pauses: your words are inserted exactly as transcribed, the Refine settings grey out with a note saying why, and your configuration is kept for when you switch back.

Voice command phrases (“new line”, “scratch that”) are English for now, whatever language you dictate in.

Refine

Refine is an optional, on-device pass that tidies what you said before it lands. There are three modes:

  • Raw — verbatim. Every word you said, nothing changed.
  • Punctuate — adds punctuation and capitalization only. No words are added or removed.
  • Clean — removes filler, fixes grammar, and keeps your meaning. Pick a casual or formal tone.

You can set a default mode per application: terminals and code editors stay on Raw, Slack, Messages, and Discord get Clean casual, Mail gets Clean formal.

Punctuate and Clean work in English and seven more languages; for the rest, Refine pauses and inserts your words as transcribed. See Languages.

Working in every app

Amanu inserts text two ways and chooses the right one for each app:

  • Accessibility insertion — places text directly at the cursor where the app supports it.
  • Paste-swap — where it does not, Amanu pastes the text and then restores your clipboard, including images and multiple items.

Either way, your clipboard is left exactly as it was.

Models

Two kinds of model, both on your disk, both replaceable.

Speech recognition

NVIDIA Parakeet TDT v3, run on the Apple Neural Engine through CoreML. One model, about 600 MB, downloaded once during onboarding. Running on the Neural Engine rather than the GPU is what keeps transcription fast and easy on the battery.

Refine

The optional cleanup models are 4-bit, run on the Apple Silicon GPU through MLX, and downloaded on demand. Raw mode uses none of them.

The default is Qwen3 1.7B (about 1 GB) — a balance of quality and speed, and the model Refine is tuned against. Settings keeps a small catalog around it: lighter and faster options when you want speed, larger and more capable ones when you want quality. Each downloads on demand.

Manage them all from Settings — download, inspect, or delete any model. They sit in Application Support like any other file, so you can always pull one again.

Privacy and history

No audio ever leaves your Mac. Transcription and Refine run on-device, in memory, while you hold the key. There is no account, no cloud, and no telemetry by default.

Amanu touches the network for exactly three things:

  • Model downloads, from Hugging Face, only when you ask for one.
  • Update checks.
  • Crash reports, only if you opt in — off by default.

Anything beyond those three would be a bug, not a setting.

History is optional. Transcripts can be kept encrypted at rest with AES-GCM, under a key that never leaves your Keychain. Choose how long to keep them — off, a number of days, or unlimited — or turn on zero-retention, which stores nothing at all. You can search, copy, delete, or clear everything.

Amanu lives in the menu bar — no Dock icon, no main window. The Nota mark animates through its states, so a glance tells you what it is doing.

9:41
  • Idle Running and ready, waiting for your hotkey.
  • Listening You are holding the key; Amanu is hearing you. The mark fills as you speak.
  • Transcribing Key released — turning speech into text on the Neural Engine.
  • Attention Something needs you: a permission to grant, or an error to read.
  • Setup Onboarding is not finished. Open the menu to pick up where you left off.

The menu gives you quick actions: turn dictation on or off, switch Refine mode, and copy a recent transcript. Launch at login is a single toggle, system-owned and also visible under System Settings > General > Login Items. Amanu follows your system default microphone, so changing input devices just works.

Settings

Open Settings from the menu bar icon. It is a single window with a sidebar:

  • General — launch at login, and checking for updates.
  • Hotkey — bind the push-to-talk key or combination, and choose hold-to-talk or toggle.
  • Microphone — pick the input device, or follow the system default.
  • Transcription — your dictation language, and the speech model: download it, check it, or remove it.
  • Refine — turn Refine on, choose Raw, Punctuate, or Clean, set the tone, manage cleanup models, and set per-app rules.
  • History — retention (off, a number of days, or unlimited), zero-retention, and search, copy, delete, or clear.
  • Privacy — crash reporting, which is off by default.

Updates

Amanu keeps itself current with signed updates, checked over the network and installed in place — no App Store, no account. Update checks are one of the three things Amanu uses the network for. You can also check on demand from Settings > General.

Your data and uninstalling

Everything Amanu stores lives on your Mac.

  • Models sit in Application Support. Remove any of them from Settings > Transcription or Settings > Refine, or delete the folder.
  • History, if you keep it, is encrypted at rest under a key in your Keychain. Clear it from Settings > History, or turn on zero-retention so nothing is ever written.
  • Settings are stored in the app’s standard macOS preferences.

To remove Amanu completely: quit it, turn off launch at login (Settings > General, or System Settings > General > Login Items), and move the app to the Trash. Delete its Application Support folder to remove downloaded models; clear history first if you want the encrypted store gone too.

Troubleshooting

Nothing was inserted

Check that Accessibility is granted under System Settings > Privacy & Security > Accessibility, and that the focused field is not a password or secure field.

The globe key does nothing, or opens the emoji picker

On some setups macOS reserves a quick tap of the globe key. Set “Press globe key to: Do Nothing” under System Settings > Keyboard, or bind a different hotkey in Amanu.

Quality is poor over Bluetooth

AirPods and similar headsets switch the microphone to a low-bandwidth mode when used as input. For the best results use the built-in microphone or a wired or USB microphone. Amanu follows your system default input, so set it there.

The audio device changed mid-dictation

Amanu rebuilds its audio pipeline when devices change. If a recording is interrupted it tells you rather than failing silently — hold the key and try again.

FAQ

What languages does it support?

Dictation works in 28 European languages — English, Spanish, French, German, Italian, Portuguese, Russian, Polish, and twenty more — chosen from a single setting. Refine’s cleanup runs for the first eight of those and pauses for the rest. See Languages.

Is it really offline?

Yes. Transcription and Refine run on your Mac. The network is used for exactly three things — model downloads, update checks, and opt-in crash reports — and nothing else.

Does it work in my app?

Almost certainly. Amanu inserts through Accessibility where an app supports it and falls back to paste-swap everywhere else, restoring your clipboard either way. The deliberate exception is password and other secure fields, which it refuses.

Will it slow down my Mac or drain the battery?

Transcription runs on the Neural Engine, which is built for this and sips power. Refine is optional and runs on the GPU only when you use it.

Where is my data?

On your Mac. See Your data and uninstalling.

How much does it cost?

Amanu is a one-time purchase of $29 with a 14-day free trial — no subscription. See pricing for what is included.